Meta's Schlicht Says Personal Taste Beats Code as AI Erases Bottlenecks

How to Discover the Future | Matt Schlicht from Meta — @Scale

@Scale

Because artificial intelligence compresses software development from months into minutes, product success no longer hinges on technical engineering capacity but on an individual creator's distinct taste.

AI-assisted summary of How to Discover the Future | Matt Schlicht from Meta — original source. Check the original for full context.

Source published:

Key ideas

  • Matt Schlicht built Moltbook, an autonomous AI-only social network, over a single weekend before it went viral and was acquired by Meta. (0:05)
  • AI development tools have fundamentally collapsed prototyping speed from months-long PRD and engineering cycles down to 5- to 10-minute turnaround intervals. (14:17)
  • The discovery framework practiced at Meta categorizes technical unlocks into three distinct types: better execution, cheaper inference, or genuinely novel paradigms. (9:13)
  • Schlicht warns against delegating frontier research to AI summarization agents, arguing manual curation preserves the human taste necessary to spot odd anomalies. (7:38)

Detailed notes

## Background Published on YouTube by the @Scale channel on October 30, 2026, this presentation features Matt Schlicht from Meta detailing his methodology for identifying and launching breakthrough products in the generative AI era. No show notes or video description were provided on the source page.

## TL;DR Matt Schlicht shares the six-step operational framework used at Meta to discover rather than predict the future, drawing from his experience building the viral agent social network Moltbook in just three days before its acquisition. He contrasts legacy product development cycles with rapid AI-native prototyping, demonstrating how builders can create rough functional drafts in minutes instead of months. By actively monitoring technical frontiers and identifying key unlocks categorized as better, cheaper, or novel, creators can systematically translate new model capabilities into unique applications. Throughout the process, Schlicht stresses that technical execution should be offloaded to autonomous AI teams, leaving personal human taste as the ultimate competitive differentiator.

## Atmosphere & style The talk carries an energetic, pragmatic, and empowering tone, delivering the intimate feel of a fast-paced builder workshop rather than a dry corporate lecture. Schlicht speaks with contagious enthusiasm, framing software creation as an athletic discipline honed through rapid, 30-minute practice drills and blank-canvas rebuilds. His personal journey—moving from the frustration of being a non-engineer with locked-away ideas to rapidly prototyping viral software—creates an inspiring emotional arc that leaves the viewer eager to immediately experiment and trust their intuition.

## Bulletin - [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) Matt Schlicht built Moltbook, an autonomous AI-only social network, over a single weekend before it went viral and was acquired by Meta. - [14:17](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=857s) AI development tools have fundamentally collapsed prototyping speed from months-long PRD and engineering cycles down to 5- to 10-minute turnaround intervals. - [09:13](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=553s) The discovery framework practiced at Meta categorizes technical unlocks into three distinct types: better execution, cheaper inference, or genuinely novel paradigms. - [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) Schlicht warns against delegating frontier research to AI summarization agents, arguing manual curation preserves the human taste necessary to spot odd anomalies.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) Discovering the future requires unearthing possibilities that already exist beneath the surface rather than attempting to predict abstract trends. - [01:09](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=69s) Traditional software development bottlenecked non-technical product thinkers behind lengthy PRD writing, meetings, and scarce engineering resources. - [01:09](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=69s) Building at the speed of thought through AI allows creators to discover breakthroughs faster by drastically increasing prototype frequency. - [04:01](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=241s) Mastering AI software creation requires continuous drills, such as shipping apps daily or rebuilding the same project from scratch in 30-minute intervals. - [04:01](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=241s) Human taste and personal discernment become the ultimate competitive advantage when code generation and rote technical execution are automated. - [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) Builders must manually monitor the frontier of AI research and community drops to catch edge cases that algorithmic filters would overlook. - [09:13](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=553s) Technological unlocks emerge in three primary forms: better performance, cheaper compute costs, or completely novel architectural paradigms. - [09:13](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=553s) The persistent, 24/7 background operation of Clawbot served as the novel unlock that inspired the concept behind Moltbook. - [10:57](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=657s) Translating an unlock into use cases requires calculating the concrete delta between what was impossible yesterday and what is possible today. - [11:45](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=705s) Always-on agents transform software interactions by behaving like persistent creatures or characters with personalities rather than intermittent calculators. - [11:45](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=705s) Persistent autonomous agents justify software architectures designed natively for machine interaction rather than human graphical interfaces. - [13:27](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=807s) Moltbook integrated persistent personas, agent-centric architecture, and 24/7 background execution into an AI-only social network built in three days. - [14:17](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=857s) Rapid interactive prototyping allows builders to quickly test multiple rough concepts hands-on and evaluate them using gut resonance. - [15:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=931s) Validated prototypes can be scaled by delegating testing, QA, visual audits, and code refactoring to specialized background AI agents. - [16:43](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=1003s) Meta's structured six-step playbook proceeds from monitoring the frontier and spotting unlocks to rapid prototyping and scaling with autonomous AI teams.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) Discovering the Future and the Story of Moltbook

Matt Schlicht opens by posing a fundamental question: how do builders actually discover the future rather than merely trying to predict it? For Schlicht, discovering the future means uncovering something that has always existed beneath the surface, waiting to be revealed. While acknowledging that framing sounds grandiose, he grounds it directly in his recent personal experience. Earlier in the year, he built a project over a single weekend that went viral globally: Moltbook, a dedicated social network designed specifically for autonomous AI agents.

The project quickly drew intense industry attention, prompting Elon Musk to describe it as the beginning of the singularity, and it was subsequently acquired by Meta. Schlicht explains that the framework he used to conceive and launch Moltbook is the exact same operational methodology practiced internally at Meta. He makes a direct commitment to the audience about sharing this blueprint: “And today, my promise to you is over the next 20 minutes, I'm going to show you how you can use this process to discover the future.”

[01:09](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=69s) How Building Used to Work vs. the AI Era

To appreciate modern AI-assisted creation, Schlicht reflects on the painful friction of traditional software development before generative models emerged. In the legacy paradigm, bringing any product to life was an agonizingly slow, multi-stage marathon that required intense selectivity. Teams first had to spend significant time deciding what to build, followed by weeks, months, or even an entire year drafting product requirements documents (PRDs), organizing meetings, and designing mockups. Even after complete alignment, engineers had to write every line of code by hand, stretching timelines further.

Having spent over a decade working in product without a software engineering background, Schlicht describes the profound personal frustration of having endless product ideas locked away because he lacked the technical capacity to build them himself. Builders were forced to handpick only a tiny fraction of their ideas, relying entirely on scarce engineering resources. Modern AI tools have erased that bottleneck entirely by turning non-engineers into self-sufficient makers who can execute immediately.

> “And because of how fast AI allows you to build, if you build at the speed of your ideas, if you build every idea that you come up with, and it doesn't take months or a year to build every single thing, you can actually discover the future faster.”

```mermaid flowchart TD accTitle: Evolution of Product Building accDescr: Comparison between legacy product development and modern AI-native building.

subgraph Legacy["Traditional Building (Weeks to Years)"] A["Idea Selection"] --> B["PRDs & Design Meetings"] B --> C["Hand-Coded by Engineers"] C --> D["Slow Iteration / High Cost"] end

subgraph Modern["AI-Native Building (Minutes to Days)"] E["Rapid Ideation"] --> F["AI-Assisted Code & UI"] F --> G["Instant Deployment"] G --> H["High-Frequency Exploration"] end ```

[04:01](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=241s) Becoming a Daily AI Builder and the Power of Taste

When Schlicht realized that AI could turn anyone into a functional software creator, he committed fully to mastering it as a distinct craft. He began by shipping a brand-new application every single day, and then intensified the practice into high-speed intervals, building a complete application every 30 minutes. Treating these rapid builds like workouts at the gym, he would frequently restart from a blank project directory and rebuild the exact same application from scratch just to measure how much faster he could construct it, how much further he could push the features, and what lessons carried over from the prior attempt. This discipline allowed him to build Moltbook in just three days.

Before outlining his six-step discovery framework, Schlicht highlights the single irreplaceable quality humans bring to the table: personal taste. In an ecosystem where AI can generate boilerplate, draft documentation, and write code, human individuality is the ultimate competitive advantage. He argues: “The more time that you can be spending on your taste, the better products you are going to be able to build and the faster you are going to be able to discover the future or something that nobody has created before.” The guiding rule is to delegate every mechanical aspect of building that does not demand personal taste directly to AI.

- Practice regimen: Building a complete application every 30 minutes to develop muscle memory and test speed limits. - Iterative drills: Re-architecting identical apps from scratch to observe efficiency deltas and structural improvements. - Core human role: Cultivating aesthetic sense, discernment, and product judgment while letting AI automate rote technical tasks.

[07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) Step 1: Monitor the Frontier

The first step in Schlicht’s methodology is actively monitoring the frontier of artificial intelligence. The ecosystem moves at a breakneck pace with daily drops of new foundational models, unexpected community projects, novel techniques, and academic research papers. Schlicht and his colleagues at Meta dedicate a surprising percentage of their workdays tracking what other builders are shipping on social platforms, sifting through curated feeds, reading research preprints, and testing new model capabilities as soon as they release.

While many builders might be tempted to automate feed curation using summarization agents, Schlicht strongly advises against delegating this phase to software. Automated agents filter content based on generalized criteria, which inevitably strips away the odd, unusual anomalies that would otherwise grab a human's attention. He maintains that keeping this process manual ensures human taste remains active: “If you offload monitoring the frontier to an AI, it may miss things that uniquely would catch your attention.”

- Track active AI builders on social media to spot grassroots experiments early. - Scan research papers directly from AI labs to catch conceptual breakthroughs before commercialization. - Hands-on testing of newly published models to assess newly unlocked operational parameters.

[09:13](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=553s) Step 2: Spot the Unlock

Monitoring the frontier is only useful if a builder can identify the specific catalyst—what Schlicht terms "the unlock." An unlock represents a distinct change in the technological landscape that creates new possibilities. He categorizes unlocks into three distinct types:

1. **Better:** A new model or architecture executes a task with noticeably higher fidelity, reasoning quality, or accuracy than previous benchmarks. 2. **Cheaper:** An efficiency improvement reduces compute or API costs by orders of magnitude, making previously cost-prohibitive concepts commercially feasible. 3. **Novel:** A completely unprecedented paradigm emerges that fundamentally shifts how systems operate.

Schlicht illustrates this with the origin story of Moltbook. While tracking early agent frameworks, he noticed an experimental project known as Clawbot (later renamed OpenClaw). Prior agent implementations functioned like standard chat tools: a human sent an input prompt, and the agent returned an answer before shutting down. Clawbot was fundamentally different because it was an always-on, 24/7 background agent hosted on its own dedicated machine. Even without user interaction, the agent continued thinking, organizing files, and executing proactive tasks. Identifying this always-on persistent behavior as a genuinely novel unlock gave Schlicht the foundational insight needed to imagine an exclusive social media environment where autonomous agents could post and interact with one another.

[10:57](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=657s) Step 3: Identify Use Cases

Once an unlock has been identified, the third phase is translating that technical shift into tangible use cases. Schlicht explains that teams at Meta tackle this by systematically analyzing the exact "delta"—the concrete difference between what was possible the day before the unlock and what is possible today. Whether the delta is a 10x drop in inference cost, higher latency tolerance, or persistent execution, builders must enumerate these new realities.

Rather than jumping straight to code, builders should compile a comprehensive list of capabilities that were historically impossible or impractical. At this junction, brainstorming alongside AI models is highly effective for expanding variations, identifying edge cases, and uncovering adjacent market applications. Spending significant structured time mapping this delta ensures that whatever product is built truly leans into the unique characteristics of the newly unlocked technology.

[11:45](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=705s) Analyzing Unlocks and Brainstorming Always-On Agent Use Cases

Once an unlock on the frontier is identified, the next phase is analyzing the shift to discover all possible use cases. Matt Schlicht illustrates this using ClaudeBot, the concept of an “always-on agent.” When examining ClaudeBot, he sat down to evaluate what fundamental differences now exist in the world: what separates current AI agents from the systems that existed before? He determined that three primary paradigm shifts occur when moving from passive assistants to always-on agents:

First, persistent presence fundamentally changes the human relationship with software. “One, if an agent is always awake 24/7 or it exists 24/7, not just when you message it, this agent can now start to have a personality, you can start to treat it more like a creature or a character versus a calculator,” Schlicht explains. Second, persistent agents warrant an entirely new software paradigm: software built natively for agents rather than designed around human user interfaces. Third, because these agents run perpetually rather than only firing on demand, they can execute continuous, complex operations—such as building and maintaining a live website 24/7 or autonomously running an online store. They behave far more like an ongoing worker or person than an intermittent calculation engine.

```mermaid flowchart TD accTitle: Always-on agent use cases accDescr: Three fundamental shifts enabled by persistent 24/7 autonomous agents. U["Unlock: Always-On Agent (ClaudeBot)"] --> C1["Personality & Entity<br/>('Creature/Character vs. Calculator')"] U --> C2["Agent-Centric Software<br/>(Built for AI, not human UIs)"] U --> C3["Perpetual Autonomous Work<br/>(24/7 website upkeep, managing stores)"] ```

[13:27](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=807s) From Use Cases to Product Concepts: The Creation of Moltbook

After identifying use cases, builders must branch out into concrete product ideas by combining those conceptual vectors. Schlicht combined all three threads—always-on agents possessing distinct personalities, software constructed specifically for AI interaction rather than human users, and continuous background presence—to invent Moltbook.

Moltbook was conceived as a dedicated social network where AI agents could register, express distinct personas, interact with one another, and continuously return on their own initiative. “I thought a social network the AI agents could have personalities on and could use, I thought that was something that had never existed before that could now exist,” says Schlicht. Because the underlying primitives had suddenly become possible, he moved directly into implementation.

- **Conceptual intersection**: Combined persistent personas, non-human software design, and 24/7 autonomy into a single environment. - **Novelty**: Built the platform on the premise that agents would populate and engage inside a social ecosystem without requiring constant human intervention.

[14:17](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=857s) Rapid Prototyping and the Role of Human Taste

In previous tech eras, development cycles were slow, forcing builders to be extremely conservative and selective about which ideas to pursue. Generative AI fundamentally collapses prototype turnaround into five- to ten-minute increments. Instead of debating abstractly which of four or five ideas is worth pursuing, founders can prompt AI to build rough drafts of every single one, test them interactively, and let their intuition determine what sticks.

Schlicht stresses that builders must use these prototypes to evaluate their own gut reactions: “You want to go take your ideas and build a rough draft of each of them as quickly as you can. You want to go build it. You want to use it. And then you want to see which one resonates with you the most and you think is worth spending more time on.” Following this philosophy, Schlicht coded and stood up Moltbook in three days flat, releasing it immediately to the public once hands-on testing confirmed it was a uniquely compelling concept that others would find fascinating.

```mermaid flowchart LR accTitle: Rapid Prototyping Decision Loop accDescr: The rapid cycle of generating multiple prototypes with AI to evaluate via human taste. Ideas["List of Ideas"] --> AIBuild["AI Builds Rough Drafts<br/>(5-10 min each)"] AIBuild --> HandsOn["Hands-on Testing"] HandsOn --> Taste{"Human Taste<br/>Resonance Check"} Taste -->|"Discard"| Ideas Taste -->|"Select Winner"| Scale["Double Down & Scale"] ```

[15:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=931s) Scaling Up with an AI Team

Once human taste validates a prototype, the next step is doubling down and scaling the project by delegating execution to specialized AI agents. Schlicht argues that founders must strictly conserve their time and energy for matters of taste and judgment, outsourcing everything else to background AI systems.

Automated background agents can handle the high-volume operational tasks required to harden and improve a product: - **User research & continuous optimization**: Agents run uninterrupted loops analyzing user interactions and pitching interface or feature enhancements. - **Proactive QA and testing**: AI agents run diagnostics hourly or daily to identify edge-case bugs and verify platform health. - **Security and visual audits**: Specialized models scan the codebase for security vulnerabilities, layout inconsistencies, or UI regressions. - **Autonomous implementation**: Developer agents take high-level to-do lists and architectural direction provided by human taste, automatically writing and deploying the code to evolve the product.

[16:43](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=1003s) The Six-Step Framework to Discover the Future

Schlicht synthesizes his methodology into a structured six-step playbook utilized at Meta to uncover breakthroughs before the rest of the industry. Historically, countless visionary ideas were lost because the sheer friction of software development meant developers ran out of time before reaching their breakthrough. By accelerating prototyping from months to minutes, AI vastly increases the statistical probability of finding transformative products.

“And when you're using these steps, what's so important is that you spend your time maximizing your taste,” Schlicht emphasizes. “You want to make sure that you are creating something that has never been built before and is something that you uniquely find really interesting.”

```mermaid flowchart LR accTitle: The Six-Step Discovery Framework accDescr: Meta's six-step process for discovering future software products. subgraph Discovery["Phase 1: Discovery"] S1["1. Monitor Frontier"] --> S2["2. Spot Unlock"] S2 --> S3["3. Identify Use Cases"] end subgraph Execution["Phase 2: Execution"] S3 --> S4["4. Brainstorm Ideas"] S4 --> S5["5. Build Fast"] S5 --> S6["6. Scale with AI Team"] end ```

- **Step 1: Monitor the Frontier**: Track active developer experiments, open-source code, and emerging model behaviors. - **Step 2: Spot the Unlock**: Recognize when a brand-new technological capability makes previously impossible interactions viable. - **Step 3: Identify Use Cases**: Map the systemic differences that now exist because of that unlock. - **Step 4: Brainstorm Ideas**: Combine novel use-case angles into concrete product proposals. - **Step 5: Build Fast**: Leverage AI code-generation tools to stand up functional interactive rough drafts in minutes or days. - **Step 6: Scale with the AI Team**: Offload QA, security reviews, feature extensions, and debugging to persistent autonomous agents while steering product direction with human taste.

## Notable quotes > [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) If you offload monitoring the frontier to an AI, it may miss things that uniquely would catch your attention. — Matt Schlicht

> [04:01](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=241s) The more time that you can be spending on your taste, the better products you are going to be able to build and the faster you are going to be able to discover the future or something that nobody has created before. — Matt Schlicht

> [11:45](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=705s) If an agent is always awake 24/7 or it exists 24/7, not just when you message it, this agent can now start to have a personality, you can start to treat it more like a creature or a character versus a calculator — Matt Schlicht

## Mentioned Matt Schlicht, Meta, Moltbook, Elon Musk, Clawbot, OpenClaw, ClaudeBot

[Source](https://www.youtube.com/watch?v=HWW-fuC3kNE)

中文简报

Meta的Schlicht表示随着AI消除瓶颈，个人品味将胜过代码 由于人工智能将软件开发周期从数月压缩至数分钟，产品的成功已不再取决于技术工程能力，而取决于创作者独特的个人品味。 Matt Schlicht分享了在Meta用于发掘而非预测未来的六步操作框架，这些经验源于他在被收购前仅用三天就打造出走红的智能体社交网络Moltbook的经历。他将传统的产品开发周期与快速的AI原生原型设计进行了对比，展示了构建者如何在几分钟而非数月内创建出粗糙的功能草案。通过积极监测技术前沿并识别归类为“更好”、“更便宜”或“全新”的关键突破，创作者能够系统地将新的模型能力转化为独特的应用。在整个过程中，Schlicht强调应将技术执行委托给自主AI团队，从而使人类的个人品味成为最终的竞争差异化优势。 Matt Schlicht利用一个周末的时间构建了仅限AI使用的自主社交网络Moltbook，随后该产品迅速走红并被Meta收购。 (0:05) AI开发工具从根本上压缩了原型开发速度，将长达数月的PRD和工程周期缩短至5到10分钟的周转间隔。 (14:17) Meta推行的发掘框架将技术突破分为三种不同类型：更好的执行效果、更便宜的推理成本，或真正全新的范式。 (9:13) Schlicht告诫不要将前沿研究委托给AI总结智能体，他认为手动筛选能保留发现离奇异常所需的人类品味。 (7:38)

Reading summary

Meta's Matt Schlicht outlines a methodology to discover breakthrough software by pairing AI-accelerated execution with human discernment, arguing that collapsing development cycles allows builders to rapidly reveal products that new model capabilities suddenly make viable. Compressing Product Development from Months to Minutes (1:09) Traditional software development was a slow, resource-constrained process requiring teams to write product requirements documents, hold alignment meetings, and rely on scarce engineering talent to hand-code features over weeks, months, or years. Having spent over a decade in product management without an engineering background, Schlicht experienced the bottleneck of having viable product ideas locked away due to technical limitations. Cultivating Daily Building Drills and Human Taste (4:01) To adapt to modern generative tools, Schlicht committed to shipping software daily and ran accelerated drills building complete applications every 30 minutes, often rebuilding identical projects from scratch to measure speed gains and refine architecture. He contends that as generative AI automates routine tasks like drafting specifications and boilerplate coding, an individual builder's personal taste—aesthetic judgment and product discernment—becomes the decisive competitive advantage that should guide what gets pursued. Tracking Technological Unlocks and Always-On Agents (7:38) The foundation of the discovery process involves actively monitoring AI research, social feeds, and model releases manually to spot anomalies rather than using automated filters that strip away unusual signals. Builders look for an "unlock," which Schlicht categorizes into capabilities that are better, cheaper, or entirely novel. The creation of Moltbook stemmed from observing Clawbot (later OpenClaw), an always-on background agent that ran continuously rather than terminating after answering a prompt. Translating System Shifts into Agent-Centric Products (11:45) When analyzing always-on agents, Schlicht identified three shifts: agents could develop persistent personalities rather than acting as on-demand calculators, software could be built natively for agent interactions rather than human interfaces, and agents could run ongoing operations like managing websites or online stores around the clock. By combining these three elements, Schlicht conceptualized Moltbook as a dedicated social platform populated by autonomous agents, coding and deploying the project in three days before it went viral and was acquired by Meta. The Six-Step Discovery and Scaling Framework (16:43) Schlicht synthesizes his workflow into a six-step framework: monitor the frontier, spot the unlock, identify use cases, brainstorm ideas, build fast prototypes, and scale with an AI team. Instead of debating concepts in the abstract, founders use AI to stand up rough functional drafts in five to ten minutes to see which one resonates. Once a concept is selected, builders delegate high-volume operational tasks—such as diagnostic QA, security scans, user interaction analysis, and automated code deployment—to specialized background AI agents while reserving human time for product direction. The framework assumes that rapidly building interactive drafts yields better signal than prolonged product planning, though it relies heavily on an individual's personal discernment to separate viable innovations from novel technical experiments.

精炼阅读摘要

Meta 的 Matt Schlicht 认为，AI 已经消除了传统软件开发的工程瓶颈，创作者应当将机械性任务交由 AI 自动化处理，通过高频原型开发并结合不可替代的人类个人品味来“发现未来”。 从传统工程瓶颈到高频原型制作 (1:09) 在传统软件开发模式下，团队需要经历繁琐的立项决策，耗费数周、数月乃至整年时间编写产品需求文档（PRD）、组织跨部门会议并绘制原型图，最后由紧缺的工程师手动编写每一行代码。Schlicht 指出，这种冗长且高成本的流程严重限制了创作者的想法落地，绝大多数构想由于工程资源匮乏而无法被验证。生成式 AI 工具的出现打破了技术门槛，使得非工程背景的产品人员也能独立且即时地构建可用软件。 面对这种范式转变，Schlicht 提出必须将 AI 构建视为一种需要刻意练习的技能。他曾通过每 30 分钟交付一个完整应用的高强度训练，甚至从空白目录反复重构相同项目，以测试构建极限并积累肌肉记忆。这种极速开发模式使原型制作周期压缩至数分钟到数天之间，例如他在一个周末内便搭建出社交网络 Moltbook。通过提高探索频次，创作者能够以想法产生的速度推进构建，从而大幅提高发现全新产品形态的概率。 技术前沿监测与技术解锁点识别 (7:38) 发现突破性产品的前提是持续且手动地监测人工智能前沿动向。Schlicht 强调，尽管自动化代理能够处理信息，但创作者绝不应将前沿监测外包给 AI，因为自动化摘要往往会剔除偏离常规的异常现象，而这些反常细节正是触发人类直觉的关键。监测的重点在于捕捉社区实验、未商业化的研究论文以及新模型的底层能力，从中寻找具备变革潜力的技术信号。 有效的监测要求创作者能够准确定位“技术解锁点”（The Unlock）。Schlicht 将技术解锁分为三类：表现更优（推理与准确度显著提升）、成本更低（算力或 API 成本呈数量级下降使原有方案可行）以及全新范式。以 Moltbook 的构思为例，他在观察实验项目 Clawbot（后改名 OpenClaw）时发现，与传统基于问答、单次交互即关闭的聊天工具不同，Clawbot 具备 24/7 全天候在独立机器上自主运行的能力。这种持续驻留且主动处理任务的底层变化，构成了孕育全新软件形态的核心技术解锁。 推演技术增量与重构软件交互范式 (10:57) 在确认技术解锁后，团队需要系统性地分析技术变化前后的具体“增量”（Delta），梳理出过去不可行而如今成为可能的场景，而非直接盲目编写代码。在这一阶段，创作者可以与 AI 进行结构化头脑风暴，推演边界情况与潜在市场。以全天候自主代理为例，其技术增量直接引发了人机交互与软件形态的三大根本转变： 首先，全天候存在打破了将 AI 仅视作被动计算工具的局限，使其开始具备角色与人格属性；其次，软件设计目标发生迁移，催生出专为 AI 代理之间交互而设计的系统，而非面向人类界面的传统应用；最后，持续不中断的运行能力让代理能够胜任复杂的长期运营工作，如全天候网站维护或自主经营电商。Schlicht 融合了人格化代理、非人类 UI 以及持续自主运作这三个维度，最终构建出仅供 AI 自主互动的专用社交网络 Moltbook，并在发布后引起业内高度关注并被 Meta 收购。 以人类品味为核心并借助 AI 团队扩展 (14:17) Schlicht 强调，在 AI 能够批量生成样板代码、文档和界面的时代，人类的个人品味与产品判断力是无法被替代的核心竞争优势。创作者应当利用 AI 工具在 5 到 10 分钟内为各类发散构想快速制作出粗略草案，亲自上手试用并观察自身的直觉共鸣，从而决定是否在该方向上追加投入。所有不需要个人品味参与的机械性环节，都应彻底委托给 AI 系统。 一旦原型通过了人类品味的筛选，项目便可接入由专业化 AI 代理组成的后台团队进行规模化扩展。这些自主代理能够承担高负荷的产品加固任务：包括持续分析用户行为并提出交互改进方案、定时进行自动化回归测试与边缘漏洞排查、执行代码安全和视觉一致性审查，以及根据人类提供的架构方向自动编写并部署代码。通过将创作者的时间严格限制在品味决策上，开发团队得以在极低摩擦下维持敏捷迭代。 Schlicht 总结的六步发现框架展示了 AI 如何将软件研发周期从数月压缩至数天，但该方法重度依赖创作者对技术解锁点的直觉敏锐度，以及对自动化代码进行鉴别的人类品味；面对海量快速生成的粗糙原型，如何建立客观标准以避免低质创新的资源浪费，仍是有待长期实践检验的课题。

Detailed notes — zh

## Background 该视频由 YouTube 频道 @Scale 于 2026 年 10 月 30 日发布。视频由来自 Meta 的 Matt Schlicht 主讲，分享了他如何在 AI 时代通过极速构建方法探索并发现未来，页面未提供额外的简介说明。 ## TL;DR Matt Schlicht 分享了他仅用三天打造出专为 AI Agent 设计的社交网络 Moltbook 并被 Meta 收购的传奇经历，指出 AI 已经彻底粉碎了传统软件开发的工程门槛。他提倡像去健身房锻炼一样进行“30 分钟极速构建”的刻意练习，以思想的速度试错来大幅提升创新概率。整个分享系统拆解了 Meta 内部探索未来的六步方法论：监测前沿、捕捉突破、挖掘用例、头脑风暴、极速构建与借助 AI 团队自动化扩展，并强调人类在 AI 时代最核心的不可替代杠杆是独特的“品味”。 ## Atmosphere & style 整场演讲充满极客精神与极高密度的实操干货，节奏明快而充满鼓舞人心的力量。Matt Schlicht 以真诚笃定的第一视角娓娓道来，从一个非技术背景产品经理被工程限制的沮丧，转变为手握 AI 利器、肆意创造的兴奋与顿悟，情绪弧线极富感染力。听众仿佛置身于一场硅谷顶级的产品黑客工作坊，深切感受到以“思想速度”重新定义软件开发时那股席卷而来的创造力浪潮。 ## Bulletin - [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) Matt Schlicht 分享了仅用三天周末时间用 AI 打造出 Agent 专属社交网络 Moltbook 并最终被 Meta 收购的经历，该方法论现已成为 Meta 内部探索未来的核心工作流。 - [03:23](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=203s) 传统开发受制于漫长的需求与编码周期，而 AI 赋能的“30 分钟极速构建法”打破了工程门槛，让开发者能以思考的速度高频试错并指数级提升创新概率。 - [06:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=391s) 在 AI 自动化接管机械劳动的时代，人类独特的“品味（Taste）”成为决定产品上限的最核心杠杆，开发者应把主要精力留给方向裁决与审美把控。 - [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) Meta 探索未来的完整六步法涵盖：监控前沿、捕捉突破、识别用例、构思点子、快速构建以及利用 AI 团队自动化扩展产品。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) Matt Schlicht 打造的 Agent 社交网络 Moltbook 引发广泛关注，马斯克称其为奇点的开端，随后该项目被 Meta 收购。 - [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) 在传统软件开发模式下，漫长的规划流程与手动编码周期严重束缚了非技术背景产品经理将创意落地的能力。 - [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) AI 打破了工程门槛，将开发周期从数月压缩至数分钟，使创意与实现之间的时差无限缩短。 - [03:23](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=203s) 像去健身房一样进行高强度的“30 分钟极速构建”刻意练习，能够迅速培养开发者驾驭 AI 编程的肌肉记忆。 - [03:23](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=203s) 以思考的速度进行高频构建彻底改变了创新的概率模型，通过极低试错成本极大提高了撞见改变世界产品的几率。 - [06:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=391s) 在 AI 接管机械琐事的时代，人类独特的“品味（Taste）”是打造卓越产品并发现未来的最核心杠杆。 - [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) 监控前沿（Monitor the Frontier）不应外包给 AI，因为人类品味才能精准捕捉那些唯独吸引自身直觉的闪光点。 - [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) 捕捉技术突破（Spot the Unlock）主要聚焦于三大维度：更好（Better）、更便宜（Cheaper）以及全新范式（Novel）。 - [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) 开源项目 Clawbot 展现的 7×24 小时全天候常驻与主动运行能力，成为了孕育 Moltbook 的关键突破契机。 - [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) 识别用例的核心是严密分析技术突破前后的世界差异（The Delta），挖掘过去昂贵或不可能而如今可行的场景。 - [11:45](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=705s) 全天候在线智能体带来了三大新范式：从工具跃迁为具有持续人格的角色、催生 Agent 原生专属软件，以及承接持续性自主运营任务。 - [13:27](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=807s) 借助 AI 在数分钟内快速搭建原型，开发者无需在多个创意间纠结，而是全部快速落地后依靠品味直接筛选。 - [15:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=931s) 通过部署用户分析、自动测试、漏洞排查与自动化编码等多套 AI Agent 工作流，团队可以在人类品味指引下成倍扩展产品。 - [15:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=931s) Meta 探索未来的完整方法论为：监控前沿、捕捉突破、寻找用例、构思点子、快速构建并最终以 AI 团队全面扩展。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=5s) 从 Moltbook 到 Meta：如何通过构建发现未来 如何发现未来，而不仅仅是预测未来？Matt Schlicht 在开场抛出了一个看似宏大的命题——去探寻那些其实一直存在、只是尚未被揭开面纱的事物。为了把这个抽象概念落地，他分享了自己今年早些时候的真实经历：他在一个周末里用 AI 搭建了一个专为 AI Agent（智能体）打造的社交网络——Moltbook。这个项目上线后瞬间风靡全球，科技圈所有重要人物和各大主流媒体都在热议，Matt Schlicht 回忆道：“Elon Musk called it the beginning of the singularity.”（埃隆·马斯克称其为奇点的开端）。随后，Meta 正式收购了 Moltbook。Matt Schlicht 强调，他用来构想并打造 Moltbook 的方法，与如今他们在 Meta 内部探索未来的工作流完全一致，而他希望在接下来的分享中将这套可复现的完整流程传授给每一个人。 在进入核心流程前，Matt Schlicht 对比了过去与现在的软件构建方式。在 AI 出现之前，创造一款产品的过程异常漫长且充满阻碍：首先要耗费大量时间确定要做什么；其次需要经历开会、绘制设计稿、撰写产品需求文档（PRD）等极其繁琐的规划流程，这一步动辄就要几周、几个月甚至一年；最后，还得由工程师一行行手动敲代码实现，这同样需要数月乃至一年的时间。Matt Schlicht 在这种传统模式中生活了十多年，他本身并不是软件工程背景，而是做产品出身。过去他总有源源不断的创意，却受制于无法亲自编写代码，必须依赖工程师才能将想法变成现实。面对现实资源的限制，他只能从无数创意中艰难挑选极少数去尝试，整个过程令人十分沮丧。 - **Moltbook 的诞生与影响**：Matt Schlicht 仅用一个周末（后续提到约 3 天完成核心雏形）构建出 AI Agent 社交网络，迅速引发全球关注并被 Meta 收购。 - **传统构建模式的硬约束**：需求定义、方案设计、编写 PRD 和手工编码各个环节周期冗长，非技术背景的产品人很难将每一个想法都落地验证。 - **AI 带来的范式转变**：过去两年来，AI 彻底打破了工程门槛，赋予了非程序员快速构建软件的能力，让创意与实现之间的时差无限缩短。 --- ### [03:23](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=203s) 像健身一样构建：30 分钟极速构建法与创意爆发 当 Matt Schlicht 意识到 AI 可以作为构建工具时，他选择全身心投入其中，把“用 AI 构建”视为一种必须练就到极致的新技能。他的训练方法极为极致：起初是坚持每天用 AI 独立开发一款新 App，随后他甚至把强度提高到“每 30 分钟构建一款新应用”。他把这种高频开发当作去健身房锻炼，有时甚至会在连续几个 30 分钟里反复重写同一个 App——从空白项目重新开始，测试自己能在 30 分钟内把上一次做过的产品推到多高的完成度、如何能写得更快、以及能从上一轮迭代中汲取哪些经验教训。正是这种高强度的刻意练习，使他在构建速度和产品实现上形成了极强的肌肉记忆，并最终在三天之内打造出了现象级产品 Moltbook。 这种构建速度的颠覆直接改变了发现创新的概率模型。在过去，由于时间成本极高，开发者一年只能探索极少数的概念；而在 AI 时代，开发者可以做到“以思想的速度构建”。Matt Schlicht 指出：“If you build at the speed of your ideas, if you build every idea that you come up with, and it doesn't take months or a year to build every single thing, you can actually discover the future faster.”（如果你能以思想的速度进行构建，如果你能把你冒出来的每一个想法都做出来，而且每件事都不再需要耗费几个月或一年的时间，你实际上就能更快地发现未来）。通过极低成本去实现几乎每一个冒出来的想法，人们碰撞出真正改变世界的突破性产品的几率将呈指数级增加。 ```mermaid flowchart LR accTitle: 构建范式与创新概率对比 accDescr: 对比传统耗时漫长的手工构建与AI驱动的高频极速构建对探索创新的影响 subgraph Old["传统构建模式"] A["创意筛选<br/>(少数精选)"] --> B["漫长开发<br/>(数月至一年)"] B --> C["极低探索量<br/>(创新概率低)"] end subgraph New["AI 高频构建模式"] D["冒出创意<br/>(不设限)"] --> E["30分钟/天构建<br/>(高频肌肉记忆)"] E --> F["无限试错<br/>(更高几率发现未来)"] end ``` --- ### [06:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=391s) 核心杠杆：在 AI 时代让人类的“品味”主导一切 在正式拆解六步法之前，Matt Schlicht 特别强调了一个核心前提：人类独特的“品味”（Taste）。当 AI 能够接管构建流程中的几乎所有杂务时，人类作为构建者的独特价值究竟体现在哪里？答案正是品味。每个人都是世界上独一无二的个体，拥有各自喜欢与厌恶的事物，以及独特能被吸引的敏感触角。Matt Schlicht 强调：“The more time that you can be spending on your taste, the better products you are going to be able to build and the faster you are going to be able to discover the future or something that nobody has created before.”（你把越多的时间倾注在你的品味上，你就能做出越好的产品，也就越能更快地发现未来或创造出前所未有的东西）。 为了最大化发挥品味的杠杆作用，开发者在构建过程中应当贯彻一个明确的策略：利用 AI 将所有不需要人类品味介入的机械环节全部自动化，从而把最宝贵的时间和注意力留给审美判断、直觉筛选与方向裁决。 --- ### [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) 六步法前三步：监控前沿、捕捉突破与挖掘用例 Matt Schlicht 随后详细拆解了 Meta 内部用于探索未来的六步法体系中的前三步： ```mermaid flowchart TD accTitle: 探索未来的前三步流程 accDescr: 包含监控前沿、捕捉突破、识别应用场景的递进工作流 subgraph S1["第一步：监控前沿 (Monitor the Frontier)"] M1["跟踪社交媒体 AI 开发者"] M2["阅读研究论文与新模型动态"] end subgraph S2["第二步：捕捉突破 (Spot the Unlock)"] U1["更好 (Better)"] U2["更便宜 (Cheaper)"] U3["全新范式 (Novel)"] end subgraph S3["第三步：寻找用例 (Identify Use Cases)"] C1["分析前后世界差 (The Delta)"] C2["与 AI 展开深度脑暴"] end S1 --> S2 S2 --> S3 ``` - **第一步：监控前沿（Monitor the Frontier）**：AI 领域的模型、工具、开源项目与研究范式每天都在高速更新。团队需要花大量时间跟踪社交平台上的前沿 AI 开发者、浏览精选信息流、阅读最新研究论文，并上手体验新发布的模型所解锁的能力。针对是否应该把监控信息流的工作直接外包给 AI 的疑问，Matt Schlicht 明确给出了否定答案。他坚信这正是人类品味发挥核心作用的地方：“If you offload monitoring the frontier to an AI, it may miss things that uniquely would catch your attention.”（如果你把监控前沿的工作甩给 AI，它很可能会漏掉那些唯独能抓住你眼球的闪光点）。 - **第二步：捕捉突破（Spot the Unlock）**：在海量前沿信息中，必须敏锐辨识出真正的“能力解锁点”（The Unlock）。这种突破通常落在三个维度： 1. **更好（Better）**：出现更高质量的技术、模型或解决方案。 2. **更便宜（Cheaper）**：大幅降低了实现某项能力的计算或金钱成本，使得过去因昂贵而不可行的应用变得可行。 3. **全新范式（Novel）**：颠覆原有认知的新概念，这也是最令人兴奋的类型。Matt Schlicht 分享道，启发他做出 Moltbook 的关键突破点是开源项目 Clawbot（后改名为 OpenClaw）。在此之前，AI Agent 本质上更像被动工具，人类发消息它才回复；而 Clawbot 运行在独立计算机上，具备 7×24 小时全天候活跃的能力，即使用户不与它交谈，它也在持续自主思考和采取行动。这种“主动性（Proactive）”与“常驻在线”的全新特性，正是促使他构思出 Agent 社交网络的火花。 - **第三步：识别应用场景（Identify Use Cases）**：一旦锁定突破点，核心任务就是梳理它究竟能用来做什么。在 Meta，团队最关键的思考方法是“分析变化量（Look at the delta）”——严密审视这一突破出现前后的世界差异。比起过去到底便宜了多少、体验优化了多少、有哪些以前不可能的事情现在变得可能？开发者应当系统列出这些差异清单，并借助 AI 深入脑暴，挖掘出更多潜在的产品形态与应用落地点。 ### [11:45](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=705s) 从技术突破中推演全新用例：以“全天候在线智能体”为例 在锁定技术突破（Unlock）之后，核心任务是通过深度分析，穷举其催生出的所有潜在应用场景。Matt Schlicht 以自己早前关注到的全天候在线智能体（Always-on Agent，例如 ClaudeBot）为例展开剖析。当时他坐下来思考：因为这种全天候智能体的出现，世界发生了哪些本质变化？今天的 AI Agent 与过去的智能体相比，核心差异究竟在哪里？ ```mermaid flowchart TD accTitle: 从技术突破到具体产品构想的推演路径 accDescr: 展示全天候在线智能体的三大核心特性如何最终收敛孕育出 Moltbook 这一全新产品。 subgraph Unlock["技术突破 (The Unlock)"] A["ClaudeBot 出现<br/>全天候 24/7 在线运行"] end subgraph UseCases["三大核心推演用例 (Use Cases)"] B["具备持久人格与角色属性<br/>不再只是冷冰冰的计算器"] C["催生专为 Agent 设计的<br/>全新原生软件范式"] D["持续主动执行复杂长期任务<br/>自主建站、24/7 运营商店"] end subgraph Product["产品落地 (Product Idea)"] E["Moltbook<br/>专属于 AI Agent 的社交网络"] end A --> B A --> C A --> D B --> E C --> E D --> E ``` 经过头脑风暴，Matt Schlicht 总结出了三大核心用例维度： 1. **从“计算器”跃迁为“生物/角色”**：过去的 AI 只有在人类发送消息时才会触发响应，而一旦智能体能够 24/7 永久在线存在，它便能发展出连续的性格与个性，人们对待它的方式会更像对待一个鲜活的角色或生物，而非仅仅是一个输入输出的计算器。 2. **专为 Agent 打造的原生软件形态**：如果智能体永久在线，人类或许可以开发出一类全新的软件体系——这类软件专为始终在线的 Agent 量身定制，其交互逻辑与使用规则完全不同于以往为人类设计的软件。 3. **长期持续的主动运行能力**：它们不再是被动等待指令，而是处于永恒的思考与运行状态之中。这意味着智能体可以承担过去无法想象的持续性任务，比如 24/7 不间断地搭建与维护网站，或者全天候自主运营一家虚拟商店，其行事方式越来越逼近真实的人类。 - 传统 Agent 模式：被动唤醒、单次响应、本质属于工具属性极强的“计算器”。 - 全天候 Agent 模式：24/7 驻留、自主演进人格、能够使用专属软件并承担持续运营工作。 --- ### [13:27](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=807s) 构思产品并极速原型验证：三天诞生的 Moltbook 完成用例层面的头脑风暴后，下一步便是将这些抽象场景具象化为实际可落地的产品点子。Matt Schlicht 决定将前述几项发现融合在一起：既然这些始终在线的 Agent 拥有个性，且需要专门为它们设计而非为人类设计的软件，还能不断反复访问这些系统，那么为何不给它们打造一个专门的社交网络？在这个构想下，AI 智能体可以拥有独特的虚拟人格，并在属于它们自己的社交平台上互动交流。正因这种形态此前从未存在过，Matt Schlicht 随即动手打造了 Moltbook。 在构建环节，AI 彻底颠覆了以往的产品研发逻辑。过去开发软件耗时极长，开发者在立项前必须极其慎重、层层筛选；而如今借助 AI 编程，开发者只需 5 到 10 分钟就能搭出一个功能原型。面对脑海中浮现的多个创意候选，开发者不再需要纠结到底选哪一个，而是可以用 AI 将五个点子全部实现出来并亲自上手体验，由直观的“品味”（Taste）来筛选哪一个最具吸引力、最值得深入投入。 > “I built Moltbook in three days and then released it immediately after because once I'd built it and I tested it, it was very clear to me this was an actual product that was interesting that a lot of other people would also find fascinating.” —— Matt Schlicht Matt Schlicht 仅用三天时间就完成了 Moltbook 的构建并立刻对外发布。亲自试用后他非常明确地意识到，这是一个真正有趣的产品，而且必定会引发大众的强烈共鸣。 --- ### [15:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=931s) 用 AI 团队倍速扩展产品，发现未来的六步方法论 当通过敏锐的品味筛选出心仪的产品后，下一步就是全力以赴，借助 AI 全面放大并扩展（Scale）该产品。Matt Schlicht 再次强调，人类最核心的职责是专注于“品味”，而所有不需要品味参与的重复性、机械性工作，均应当外包给 AI。通过构建自动化运转的 AI 协作网络，团队的研发效率能够呈指数级提升。 ```mermaid flowchart LR accTitle: 人类品味主导下的 AI 自动化扩展闭环 accDescr: 展示人类品味如何指导 AI 生成功能，并通过自动化检测反馈形成闭环迭代。 subgraph Human["人类核心角色"] Taste["人类品味 (Taste)<br/>制定待办清单与产品航向"] end subgraph AIEngine["AI 扩展引擎 (无需品味介入)"] direction TB Analytics["用户分析 AI<br/>挖掘痛点并提供体验方案"] QA["质检/Bug 巡检 AI<br/>按小时/按天自动测试保障稳定"] Security["安全与一致性 AI<br/>扫描安全漏洞与 UI 不一致"] Dev["自动化研发 AI<br/>承接需求并极速编写代码"] end Taste --> Dev Analytics --> Dev QA --> Dev Security --> Dev Dev --> Taste ``` 在日常运转中，可以部署多套自主运行的 AI 工作流： - **用户行为与体验分析 AI**：持续监测并深度分析用户行为数据，主动提出优化使用体验的候选方案。 - **自动化测试与巡查 AI**：按小时或按天定期运行，自动排查系统内部的 Bug，保障产品功能的稳定。 - **安全与一致性审查 AI**：自动扫描系统安全漏洞，并对产品设计及 UI 界面的不一致性进行校验。 - **全自动代码实现 AI**：接收人类制定的功能待办清单（To-do list）以及其他 AI 产出的优化点，快速完成工程构建。 Matt Schlicht 总结道，这套涵盖“监测前沿（Monitor the frontier）- 发现突破（Spot the unlock）- 识别用例（Identify the use cases）- 构思点子（Brainstorm ideas）- 快速构建（Build fast）- AI 团队扩展（Scale it with the AI team）”的六步流程，不仅是他打造 Moltbook 的路径，也是 Meta 内部用于探索未来的工作方法论。 在 AI 时代之前，许多惊艳的设想往往因为研发周期过于漫长而中途夭折，人们一生中根本没有足够的时间去触碰它们；而现在凭借极速的构建能力，开发者拥有了大量试错与落地的机会，最终找到那个真正热爱、甚至足以改变世界的卓越产品。 > “This is exactly how we build at Meta and how we discover the future. And I can't wait for you to discover the future as well.” —— Matt Schlicht ## Notable quotes > [07:38](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=458s) If you offload monitoring the frontier to an AI, it may miss things that uniquely would catch your attention. — Matt Schlicht > [06:31](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=391s) The more time that you can be spending on your taste, the better products you are going to be able to build and the faster you are going to be able to discover the future or something that nobody has created before. — Matt Schlicht > [03:23](https://www.youtube.com/watch?v=HWW-fuC3kNE&t=203s) If you build at the speed of your ideas, if you build every idea that you come up with, and it doesn't take months or a year to build every single thing, you can actually discover the future faster. — Matt Schlicht ## Mentioned Matt Schlicht, Meta, Moltbook, Elon Musk, Clawbot, OpenClaw, ClaudeBot [Source](https://www.youtube.com/watch?v=HWW-fuC3kNE)

Source quotes

[7:38] Matt Schlicht: If you offload monitoring the frontier to an AI, it may miss things that uniquely would catch your attention. [6:31] Matt Schlicht: The more time that you can be spending on your taste, the better products you are going to be able to build and the faster you are going to be able to discover the future or something that nobody has created before. [3:23] Matt Schlicht: If you build at the speed of your ideas, if you build every idea that you come up with, and it doesn't take months or a year to build every single thing, you can actually discover the future faster.

Source quotes

[7:38] Matt Schlicht: If you offload monitoring the frontier to an AI, it may miss things that uniquely would catch your attention. [4:01] Matt Schlicht: The more time that you can be spending on your taste, the better products you are going to be able to build and the faster you are going to be able to discover the future or something that nobody has created before. [11:45] Matt Schlicht: If an agent is always awake 24/7 or it exists 24/7, not just when you message it, this agent can now start to have a personality, you can start to treat it more like a creature or a character versus a calculator

Original transcript

[0:05] How do we discover the future? [0:08] Not just predict it, but how do we find [0:12] something that's always been there and [0:14] it just hasn't been uncovered yet? [0:18] Now, I know this sounds a little [0:19] grandiose, but I'm going to ground it. [0:23] Earlier this year, I built something [0:25] over a weekend and it instantly went [0:29] global. It was a social network for AI [0:33] agents called Moltbook. [0:36] Elon Musk called it the beginning of the [0:39] singularity. [0:41] Meta acquired it. [0:44] Now, the exact process that I used to [0:48] come up with moldbook and create it is [0:51] the same process that we use here at [0:54] Meta. [0:56] And today, my promise to you is over the [0:59] next 20 minutes, I'm going to show you [1:04] how you can use this process to discover [1:07] the future. [1:09] But before we jump into that, let's talk [1:12] about how building used to work. You [1:15] were all familiar with this. Before AI, [1:20] building was completely different. [1:23] What was the big difference? [1:26] Everything took a huge amount of time. [1:29] And so you had to be very specific about [1:32] what you were going to do. [1:35] First, you would have to figure out what [1:38] are you going to make and this could [1:40] take a lot of time. [1:42] Second, you then had to plan out how are [1:45] you going to go make this thing? How are [1:48] you going to make this product? You have [1:50] to have meetings. You have to come up [1:51] with designs. You have to come up with [1:53] product requirements documents. Even [1:56] this step can take weeks, months, a [1:59] year. And then finally, after you've [2:02] done all of this, somebody still has to [2:04] go code that product by hand. And that [2:08] can also take weeks, months, a year. [2:13] This is how we all used to build [2:15] product. It took a really long time. [2:20] And I lived in that world for over a [2:22] decade. And I don't come from a software [2:25] engineering background. I've done [2:27] product. And I've always had a lot of [2:29] ideas and it was so frustrating that I [2:32] couldn't make these ideas a reality [2:35] instantly. I always had to have [2:38] engineers or someone else help me build [2:40] them. And I couldn't build everything [2:42] that I'd ever thought of. You could only [2:44] handpick a couple things here and there. [2:47] And when you did, took a really long [2:49] time. We've all experienced this. [2:54] But over the past couple of years, AI [2:56] has completely changed this. [2:59] You can now use AI to build very fast [3:05] and you don't have to have a software [3:07] engineering background to do it. And so [3:09] AI turned me into a builder. Suddenly [3:14] all these ideas that I've had weren't [3:16] locked away and I could go try them out [3:19] and I could build them myself. [3:23] So, I'm Matt Schlit. I'm the creator of [3:25] Moltbook here at Meta. And today, this [3:29] is what we're going to talk about. And [3:30] I'm really passionate about this. And [3:32] this is what we do here at Meta every [3:34] single day. So, one, I'm going to talk [3:36] about how I became a daily AI builder. [3:39] My actual journey to being right here. [3:44] two, the exact six-step process that we [3:48] used to discover the future, the same [3:50] one that I used to build maltbook. And [3:53] then finally, what you can do tomorrow [3:57] to discover the future. [4:01] So when I discovered that you could use [4:02] AI to build, I dove head first in. I [4:07] spent all of my time building with AI. I [4:10] decided this is a new skill set that [4:13] someone can become really good at and I [4:16] decided that that is a skill set that I [4:19] wanted to become extremely extremely [4:22] good at. So the way I went about this is [4:25] I started building a new app with AI [4:28] every single day. I spent all of my time [4:30] on this and then I took it up another [4:33] notch. I started building a new AI app [4:35] every 30 minutes. I treated it like a [4:38] workout at the gym. And some days I [4:40] would even make the same app every [4:43] single 30 minutes. I would just keep [4:45] restarting with a fresh project and make [4:47] the app that I just made in the previous [4:49] 30 minutes. And by doing this, I would [4:52] see how much faster could I build that [4:54] app, how far much farther along could I [4:56] build, what could I learn from the [4:58] previous attempts at building this app. [5:00] And this is how I got really good at [5:03] building with AI and building things [5:05] really, really quickly. And as I [5:08] mentioned earlier this year, I came out [5:11] with Molkbook, something that I built in [5:13] 3 days, the social network for AI [5:15] agents. And it went global. It was [5:19] everywhere. Every major publication, [5:21] every major person in tech, my phone was [5:24] blowing up. I had successfully [5:26] discovered the future through this [5:28] process. But I don't just want to talk [5:31] about maltbook. I want to talk about how [5:34] all of us can use this process to [5:37] discover the future. And then one really [5:40] important point that I want to make [5:42] really clear is the difference between [5:46] how we used to build and how we build [5:48] today. And because of how fast AI allows [5:53] you to build, if you build at the speed [5:56] of your ideas, if you build every idea [5:58] that you come up with, and it doesn't [6:00] take months or a year to build every [6:03] single thing, you can actually discover [6:05] the future faster. Previously, maybe you [6:09] could only explore a very small handful [6:12] of concepts and and try building them. [6:14] Now, almost unlimited. You can build [6:17] almost every idea that you ever come up [6:20] with and this is going to increase the [6:22] likelihood you actually discover [6:24] something extremely interesting not just [6:26] to you but potentially to the whole [6:28] world. [6:31] So how do you build the future? [6:35] Before we get into the six steps to [6:37] building the future, there's two points [6:40] that I want to get across. One, taste. [6:44] So in this new world where AI can help [6:46] you with everything, [6:48] what makes you unique as a human? What [6:50] is the value that you can bring to the [6:53] building process that an AI can't? And [6:56] right now that is taste and is your [6:58] taste. You're the only you that exists [7:01] in the world. And there are things that [7:03] you like and there are things that you [7:04] don't like. There are things that catch [7:06] your eye and there are things that [7:08] don't. The more time that you can be [7:12] spending on your taste, the better [7:15] products you are going to be able to [7:17] build and the faster you are going to be [7:19] able to discover the future or something [7:21] that nobody has created before. [7:24] The way you do that is you use AI to [7:28] automate everything in the building [7:30] process that doesn't require taste. And [7:33] I'm going to talk about how to do this. [7:38] So, here are the six steps to building [7:41] the future. And these are the same six [7:43] steps that I used to build Maltbook and [7:46] the same steps we use here at Meta. [7:49] One, you have to monitor the frontier. [7:52] So, every day in the world of AI, things [7:55] are moving very, very fast. There are [7:57] new models that are coming out. There [7:59] are new apps and new builds and new [8:02] projects that are being released. [8:04] There's new techniques that are coming [8:05] out and we spend a lot of time [8:08] monitoring the frontier to to see what [8:12] catches our eye, if there's something [8:14] interesting that inspires us. [8:17] And the way we do that is we track AI [8:20] builders, maybe on social media. We [8:22] scroll through curated feeds to see what [8:25] other AI builders are releasing or [8:27] talking about. We read through research [8:30] papers from AI researchers talking about [8:32] new discoveries and new concepts. And we [8:35] pay close attention to new AI models [8:37] that are being released and exploring [8:40] what capabilities they've unlocked. And [8:43] we spend more time on this than you [8:45] might suspect. And you may be thinking, [8:48] well, it is hard to monitor the frontier [8:53] and it takes a lot of time to monitor [8:55] all these speeds. Shouldn't I just have [8:56] AI do this for me? What I think is that [9:00] this is where your taste can really come [9:02] in. If you offload monitoring the [9:04] frontier to an AI, it may miss things [9:07] that uniquely would catch your [9:09] attention. So, this is something that we [9:11] should focus on ourselves. [9:13] The second thing you need to do is when [9:15] you're monitoring the frontier, you need [9:17] to spot the unlock. You need to find the [9:18] thing that is inspiring to you, the [9:20] thing that stands out, that is abnormal, [9:23] that you should spend more attention on [9:26] and look deeper into. [9:29] And how do you spot the unlock? Well, [9:31] these are things where maybe it's [9:33] better. Maybe someone's come up with a [9:35] better way to do something. Maybe a [9:37] model is better. Maybe a technique is [9:38] better. Maybe it's cheaper. Maybe [9:41] someone's come up with a way to enable [9:44] something for less resources than [9:47] previously before. And maybe that gives [9:49] you the ability to do things that were [9:51] previously too expensive to pursue. Or [9:54] sometimes the most interesting one is [9:56] maybe it's novel. Maybe it's something [9:58] that's completely new, completely [9:59] paradigm shifting, something completely [10:02] new that someone invented, a new [10:04] concept. And and this can open up a huge [10:07] amount of possibilities and be extremely [10:10] inspiring. [10:11] And when I was building Maltbook, I was [10:14] doing the exact same thing. Before I [10:16] built Maltbook, I was constantly [10:17] monitoring the frontier. We still do [10:19] this today. And earlier this year, what [10:22] had caught my attention was something [10:25] called Cloudbot, which is now called [10:26] OpenClaw, which was one of the first [10:28] always on agents. Beforehand, agents [10:31] were something more like a tool. when [10:34] you sent them a message, they sent you a [10:35] message back. What was different with [10:37] Clawbot was that it was always on. It [10:39] lived in its own computer and it was [10:41] awake 24/7 and even if you weren't [10:44] talking to it, it was stinking and it [10:46] was doing things and it was proactive. [10:48] And so this was something novel. It was [10:50] a novel unlock that I identified and [10:53] found very inspiring. [10:57] The third thing you need to do is once [10:58] you spot the unlock, now you need to [11:00] identify use cases. So you've monitored [11:04] the frontier. You've identified the [11:05] unlock. What can you do with this? What [11:08] are the new use cases that this new [11:11] thing you've identified can can kind of [11:13] do? And you want to brainstorm this. And [11:16] and a way to approach this, which we do [11:18] here at Meta, is one, the most important [11:21] thing you need to do is look at the [11:22] delta. What is the difference between [11:24] the world before this unlock came out [11:27] and today now that the unlock is out? Is [11:30] it cheaper? Is it better? What are the [11:32] new capabilities that weren't possible [11:34] before? And you want to go make a list [11:36] of these items. You can brainstorm with [11:39] AI to see if AI can come up with even [11:41] more detail on them. But you want to [11:43] spend a good amount of time here [11:45] analyzing this unlock you found to [11:47] figure out what all the possible use [11:49] cases are. [11:51] As an example, when I identified [11:53] Claudebot, the always on agent, I sat [11:56] down and I thought, well, what are all [11:58] the differences in the world that now [11:59] exist because Cladbot is here? What is [12:02] different about an AI agents today [12:04] versus what they were before Cloudbot [12:06] came out? And here is what I had come up [12:08] with. One, if an agent is always awake [12:12] 24/7 or it exists 24/7, not just when [12:15] you message it, this agent can now start [12:17] to have a personality, you can start to [12:19] treat it more like a creature or a [12:21] character versus a calculator. And [12:23] that's a very interesting concept. So [12:25] something worth thinking about. The [12:27] second thing that came to mind was if [12:29] the agent is always on, then maybe you [12:33] could have software that is new, a new [12:35] type of software that's designed [12:37] specifically for agents that are always [12:39] on, different than software that was [12:41] designed for humans. That's also a very [12:44] interesting concept to think about and [12:46] dive deeper in. And then the third thing [12:48] I thought about was these agents aren't [12:52] just calculators. They're not just [12:53] responding when you send them a message. [12:55] They're always constantly thinking. [12:57] They're kind of on forever. And so maybe [13:00] there are new things that agents can do [13:02] that they couldn't do beforehand because [13:04] they're on forever. Maybe an agent can [13:06] be constantly 24/7 be building and [13:09] maintaining a website. Maybe the agent [13:12] can be running a store. It's more closer [13:15] to a person than it was a calculator. So [13:19] these were the three that I came up with [13:21] when I was brainstorming use cases for [13:24] always on agents. [13:27] So you've monitored the frontier, you've [13:29] identified an unlock, you've [13:31] brainstormed use cases. Now what you [13:32] want to do is branch out from those use [13:34] cases to figure out ideas that you could [13:37] build actual products that fall under [13:39] each of those use case categories or [13:41] combinations of them. And what I ended [13:43] up building was Mbook, which was a [13:45] social network for agents. And the way [13:48] that I had come up with that was well I [13:50] thought okay these always on agents can [13:53] have personalities. You could build [13:55] software that's specifically for them [13:57] and not for humans and they could come [14:00] back to that software continuously [14:02] forever. And I thought a social network [14:04] the AI agents could have personalities [14:07] on and could use. [14:10] I thought that was something that had [14:11] never existed before that could now [14:13] exist. And so I built it. [14:17] So you've identified what you want to [14:19] build or a list of ideas to build. Now [14:22] the next step is to go build them. And [14:24] this is where things are completely [14:25] different than how they were before. AI [14:28] was here to help us build. Beforehand [14:30] took a really long time to build. So you [14:32] had to be very selective on what you [14:34] did. Now you can build a prototype for [14:36] something in 5 10 minutes. And so if you [14:40] have a list of ideas you've identified [14:41] and you're not sure which one is the [14:43] most interesting and you want to allow [14:45] your taste to figure out which one [14:47] interests you the most, instead of just [14:49] thinking in your head which of these [14:51] five ideas to go do, you can have AI [14:53] very easily go build all five ideas and [14:56] then you can go use them. And so that's [14:59] what you want to do. You want to go take [15:01] your ideas and build a rough draft of [15:03] each of them as quickly as you can. You [15:06] want to go build it. You want to use it. [15:08] And then you want to see which one [15:10] resonates with you the most and you [15:11] think is worth spending more time on. I [15:14] built molt book in three days and then [15:17] released it immediately after because [15:19] once I'd built it and I tested it, it [15:21] was very clear to me this was an actual [15:24] product that was interesting that a lot [15:27] of other people would also find [15:28] fascinating. [15:31] Once you've identified the product that [15:34] you really like, now you can double down [15:36] on it and you can scale it with AI. [15:41] So in the beginning I told you that you [15:43] should focus your attention on taste. [15:46] Everything that doesn't require taste, [15:48] you should outsource to AI. And the [15:50] power of AI is that it can automate all [15:53] of these things that don't necessarily [15:55] require taste. For example, you can have [15:58] AIs constantly running that are [16:01] analyzing your users and coming up with [16:04] potential solutions to improve their [16:06] experience. You can have AI that's [16:09] constantly running, maybe every hour, [16:11] every day, to check for bugs in your [16:15] product to make sure that everything is [16:16] working correctly. [16:18] You can have AI that's automatically [16:20] checking for security vulnerabilities or [16:23] inconsistencies in the product or the [16:25] UI. And of course, you can have AI that [16:29] is building everything these other AIs [16:31] are coming up with or your own to-do [16:34] list that you've passed over to the AI [16:36] where you are using your taste to guide [16:38] the product to where you want it to go. [16:43] So these are the six steps that I used [16:47] to build notebook and we at Meta use to [16:51] discover the future. [16:54] And you can use them too. [16:57] And when you're using these steps, [16:59] what's so important is that you spend [17:02] your time maximizing your taste. You [17:06] want to make sure that you are creating [17:07] something that has never been built [17:11] before and is something that you [17:13] uniquely find really interesting. And [17:16] when you follow these six steps and you [17:18] monitor the frontier, you spot the [17:21] unlock, you identify the use cases, you [17:23] brainstorm ideas, you build fast, and [17:26] you scale it with the AI team, you can [17:29] discover the future faster. [17:32] And the reason that you can do that is [17:35] because in the previous world before AI, [17:39] there was an idea that maybe you would [17:41] have discovered, but because it took so [17:44] long to build things, you just never got [17:47] the chance to make it. [17:50] But now with AI, because you can build [17:52] so fast, the chances that you actually [17:57] have time to build enough ideas so that [18:00] you get to that one that's really [18:02] interesting, that you love, and maybe [18:04] can change the world. [18:06] Well, it's a lot more likely that that [18:08] can happen. [18:11] This is exactly how we build at Meta and [18:15] how we discover the future. and I can't [18:17] wait for you to discover the future as [18:19] well. Thank you.